中文精读
判断 AI 机会，先分清你在押注哪一层能力
嘉宾：汪华，创新工场合伙人（当期身份）
主持：曲凯
原发布：2024 年 8 月 11 日｜时长：45 分 16 秒
整理：2026 年 10 月 18 日｜官方单集
一种常见的 AI 创业讨论，是把三个问题混成一个：语言交互会不会改变软件？机器能否独立做完复杂工作？通用人工智能何时到来？于是，一次模型发布就足以让人同时上调或下调对整个产业的判断。
汪华在 2024 年这次对谈里提供的价值，是把这些判断拆开。不同能力需要不同证据，不同应用又受不同成本约束。移动互联网可以提供参照，但不能替 AI 交出答案。理解这点，才能把“相信长期趋势”翻译成今天具体能做、能观察、能承受的事情。
一、三个阶段，是三组条件，不是一张必然兑现的时间表
回听：09:18—17:25。 汪华划分的第一层，是语言与多模态带来的交互变化，以及有人参与的简单任务自动化。用户不再必须学会每个菜单，也能用文字、声音或图像表达需求。人仍负责判断、接力和收尾。他把这层变化类比移动互联网，对其发生表达了很强信心。
第二层是复杂、长流程的通用自动化：机器接下目标，在较少人工介入的情况下连续处理多个环节。这里除了更强模型，还需要工程系统可靠地组织执行。他用蒸汽机作类比，并给出比第一层低得多的主观把握。第三层才是通用人工智能，边界和到达方式都更不确定。他说出的“百分之百”“百分之五十”是个人判断，不是经过校准的概率或行业统计。
这一区分直接影响产品论证。让系统起草一封邮件，用户检查后发送，可以证明交互和局部生产力有价值；它不能证明系统能自己协调一整次跨部门项目。后者还要持续记住状态、处理异常、确认权限并承担失败成本。
编辑推论： 路演中一句“模型已经能做”，应该被拆成“在哪个输入范围、谁补齐哪一步、错误怎样被发现”。拿第一层演示去支持第二层商业承诺，是能力与责任之间的跳步。
二、为什么机会不会同时落在所有行业
回听：01:28—08:58、18:50—20:57。 回顾移动互联网，汪华提到几种误判：用当时昂贵的设备和流量价格推算永远没有市场；把运营商时代的增值业务直接搬过去；或因为最早一批产品没跑通，就认为新平台没有独立机会。短视频也没有简单复制早期 PC 视频的产品形态。
但 AI 与移动互联网有一个重要差异。移动互联网可以增加使用场景、相机入口和联网人群，很多底层网络技术已经相对稳定。AI 的技术栈和能力边界仍在演进，也未必拥有同样规模的新硬件普及红利。因此，不能只按“上一轮哪个 App 最先成功”来排这一轮的机会。
汪华当时提出的顺序，是高价值、能覆盖推理成本的企业生产力需求先走，随后是大用户工具、图文社交娱乐，再到成本更高的多媒体和更深的业务改造。这是有前提的判断：低频但高价值的任务，可能容忍较高的单次成本；长时间陪伴或大量生成视频的产品，则可能被持续调用的账单卡住。
成本不是唯一变量。一个生成结果便宜但频繁需要返工的系统，仍可能很贵；一个调用贵但能减少大量人工交接的系统，也可能成立。要比较的是完成同一件事的总成本，而不是只比一次模型请求的价格。
三、先发优势要能带走，巨头优势也有适用范围
回听：21:23—28:32、35:43—41:32。 如果新能力只是替已有需求换一种前端交互，已有公司可以利用用户、渠道和现成业务流程。用户原本就在这里完成任务，新增一个智能入口就能接住流量。创业团队并不会因为“AI 原生”四个字自动获得优势。
当新形态要求改变收入来源、重写组织方式，甚至与旧业务相冲突时，新公司才可能拥有更大的空间。这里的重点也不是巨头一定反应慢，而是同一家公司要在保护既有业务、投入最大战场和探索边缘需求之间分配资源。一块市场被忽略，可能提供入口，但不意味着版权、运营或交付负担消失。
对先行者，汪华也没有给出“越早越好”的结论。技术变化太快，今天的形态可能不是最终形态。团队需要转向，而用户、数据、资金和积累的能力能否帮助转向，决定早期投入留下什么。
编辑推论： 评估先发优势时，可以追问：“如果六个月后产品界面必须重做，哪三样东西还属于你？”答案若只是旧界面的代码，优势很薄；若是愿意持续合作的客户、可合法使用的反馈和可靠的交付能力，才可能迁移。这里的数据也不能默认包含一切用户内容，使用权和适用范围必须另行确认。
四、开源还是闭源，应该算完整的运行账
回听：28:43—34:34。 汪华曾预测，到 2025 年上半年，GPT-4 级别能力的使用成本可能降到每个日活用户几分钱人民币。他在节目中以用户消耗 token 的假设推演。token 是模型处理文本的计量单位，但这段讨论没有统一输入输出比例、质量、延迟和重试条件。
因此，这是一项 2024 年的预测，不能直接当作今天的采购报价，更不能据此承诺所有产品都能盈利。即使模型单价下降，使用频次、上下文长度、工具调用和人工复核也可能扩大总账单。
他谈开源与闭源时同样强调阶段：小团队可能更适合使用托管 API 或第三方服务，减少维护负担；有能力的中等团队可以考虑自部署；特别大规模的应用可能混合多类模型，通过定制控制成本和差异。这是一种组织能力与经济性的匹配思路，不是按人数强行分档。
自部署并不等于免费。工程维护、扩缩容、故障处理、许可证限制和数据要求都要计入；托管服务也不等于万事交给供应商，产品仍须决定哪些内容可以发送、怎样记录失败。选型的目标是稳定完成任务，而不是选一个看起来更有立场的阵营。
五、从用户最终意图出发，也要给验证留足时间
回听：32:46—45:05。 汪华把搜索理解为很多任务共有的中间步骤：人真正想要的是完成研究、处理具体事务，搜索只是为了凑齐所需信息。如果 AI 能围绕最终意图组织服务，未来的产品就未必只是给旧 App 的每个按钮加一个聊天入口。
但“理解意图”与“完成服务”之间仍有距离。生成方案、找到信息、取得用户同意、调用服务、确认完成，不能被一条自然语言答复掩盖。第一节的能力分层在这里重新出现：交互可以先变化，后台交付未必同时自动化。
节目后段谈资本与认知。汪华认为，风险投资样本少、结果反馈慢；如果资金过分要求可预测的短期盈利，就会减少对不确定探索的支持。这是在说明一种投资机制的张力，不应读成所有某类基金已经消失，也不构成个人资产配置建议。
对产品团队，可借用的是节奏管理：不要每逢发布会就重写全部判断，也不要以长期主义为由拒绝看用户。长周期判断与短周期实验可以同时存在——长期假设决定探索方向，短期证据决定哪条路径继续投入。
应用练习：把一个 AI 方案拆成三层证据
选一个现有项目，写下用户从提出需求到确认完成的全过程，再逐项填写：
| 方案中的主张 | 至少需要什么证据 | 目前还不能推出什么 |
|---|---|---|
| 用户更容易表达需求 | 真实任务中减少理解和操作负担 | 已经能无人执行全部任务 |
| 某个步骤能自动完成 | 代表性输入、失败样本和人工接手记录 | 跨步骤长期运行同样可靠 |
| 整段流程可以交付 | 状态、权限、异常恢复、成本与完成确认 | 能无限扩展到任何行业 |
| 先发投入有长期价值 | 用户关系、合法反馈和能力可以迁移 | 早进入必然赢过后来者 |
最后另写两行：当前经济性依赖哪条成本假设；如果该假设晚一年兑现，团队会缩小什么范围。这比只争论 AI 是否“还值得相信”更接近可执行的选择。
来源边界
本文依据本期官方介绍与逐字稿整理。产业路径、成本下降和资本判断均属于 2024 年对谈中的观点，未独立验证其后兑现情况。约 18 分钟一段转写有明显否定词与上下文冲突，未据此下精确结论；不确定的设备价格和模型指标也未采用。编辑推论和练习不冒充嘉宾原话。
回听导航
- 01:28—08:58：移动互联网早期误判与 AI 的差异。
- 09:18—17:25：三层能力及其不同条件。
- 18:50—23:23：机会顺序与既有公司的优势。
- 23:51—28:32：创始人适应力与先发投入。
- 28:43—34:34：历史成本预测、开闭源与最终意图。
- 35:43—45:05：巨头的边界、资本机制与反馈周期。
相关主题：AI 产业阶段、产品经济性、创业时机。延伸收听：别刚起跑就想冲刺、Grasp 与智能体执行环境。
English edition
Before judging an AI opportunity, identify the capability you are betting on
Guest: Wang Hua, partner at 创新工场, his role at the time
Host: Qu Kai
Originally published: August 11, 2024 | Duration: 45 minutes 16 seconds
Edited: October 18, 2026 | Official episode
AI startup discussions often collapse three questions into one: will language change software interfaces, can machines independently finish complex work, and when will artificial general intelligence arrive? A single model launch can then change someone's view of the whole industry.
Wang Hua's useful contribution in this 2024 conversation is to separate those questions. Different capabilities require different evidence, and different applications face different cost constraints. Mobile internet offers comparisons, but cannot supply AI's answers. This distinction makes a long term belief more useful for deciding what to build, observe, and afford today.
1. Three stages describe conditions, not a guaranteed timetable
Listen: 09:18–17:25. Wang's first layer covers language and multimodal interfaces, together with simple tasks that still involve people. Users can express needs through text, voice, or images without learning every menu. People remain responsible for judgment, handoffs, and completion. Wang compares this change to mobile internet and expresses strong confidence in it.
The second layer is general automation of complex, extended workflows: a machine accepts a goal and handles multiple stages with less human involvement. This requires stronger models and engineering systems that organize execution reliably. He compares it to the steam engine and expresses considerably less certainty. The third layer is artificial general intelligence, whose boundaries and route remain more uncertain. His references to “100 percent” and “50 percent” are personal judgments, not calibrated probabilities or industry statistics.
The distinction changes a product argument. Drafting an email that a person checks and sends can demonstrate useful interaction and local productivity. It does not demonstrate autonomous coordination of an entire project across departments. That would also require maintained state, exception handling, permission checks, and an acceptable cost of failure.
Editorial inference: “The model can do it” should become “under which inputs, with whom completing which steps, and how are mistakes discovered?” Using a first layer demo to support a second layer commercial promise skips the gap between capability and responsibility.
2. Opportunities will not arrive in every industry at once
Listen: 01:28–08:58 and 18:50–20:57. Wang recalls several mistakes from mobile internet: projecting expensive devices and data plans indefinitely; importing the old carrier value added services model; and concluding that a new platform has no independent opportunities because its earliest products failed. Short video did not simply copy the form of early PC video products.
But AI differs in a crucial respect. Mobile internet added usage settings, camera entry points, and connected users on top of comparatively stable network technology. AI's technical stack and capability boundaries are still evolving, and it may not receive a comparable dividend from new hardware adoption. The order in which yesterday's apps succeeded cannot determine today's opportunities.
Wang's proposed sequence starts with valuable business productivity tasks that can cover inference costs, followed by tools with large audiences, image and text social entertainment, and eventually more expensive multimedia and deeper business restructuring. This is conditional reasoning. An infrequent but valuable task may tolerate a high cost per execution. Long entertainment sessions or extensive video generation may be constrained by recurring usage bills.
Cost is not the only variable. A cheap output that repeatedly needs repair can still be expensive. A costly invocation that removes substantial manual coordination may make economic sense. The relevant comparison is the total cost of completing the same job, rather than the price of one model request.
3. Early advantages must survive change; incumbent advantages have limits
Listen: 21:23–28:32 and 35:43–41:32. When a new capability merely changes the interface for an existing need, established companies can use their users, distribution, and operating workflows. Customers already complete the task there, so an intelligent entry point can capture demand. Calling a startup “AI native” does not automatically confer an advantage.
New companies may find more room when the new form changes revenue sources, requires organizational restructuring, or conflicts with the old business. This does not mean incumbents must respond slowly. An established company divides resources among protecting existing revenue, fighting its largest battles, and exploring marginal needs. A neglected market may offer an opening without removing copyright, operational, or delivery burdens.
Wang also avoids concluding that entering earlier is always better. Technology may change enough that today's form is not the eventual form. Teams must adapt. Whether users, data, capital, and accumulated capabilities support that transition determines what the early investment leaves behind.
Editorial inference: Ask, “If we must rebuild the interface in six months, which three assets remain ours?” Old interface code alone is a thin answer. Continuing customer relationships, feedback that can legally be used, and reliable delivery capabilities may transfer. Data rights cannot be assumed to include every piece of user content; permission and permitted uses need separate checks.
4. Open or closed models require a complete operating calculation
Listen: 28:43–34:34. Wang predicted that by the first half of 2025, using GPT-4 level capabilities might cost only a few renminbi cents per daily active user. His discussion used assumptions about token consumption. Tokens measure the text processed by a model, but the conversation did not standardize input versus output, quality, latency, or retries.
This is a forecast made in 2024, not a current procurement quote or a promise that every product will become profitable. Lower unit prices can coexist with higher total bills as usage frequency, context length, tool calls, and human review expand.
His discussion of open and closed models likewise emphasizes stages. A small team might use hosted APIs or third party services to reduce maintenance. A capable medium sized team might consider its own deployment. Very large applications might mix and customize models to manage costs and differentiation. This matches organizational capability to economics; it is not a rigid classification by headcount.
Operating an open model is not free. Maintenance, scaling, incident response, licensing, and data requirements belong in the calculation. Hosted services do not remove the product team's responsibility to decide what may be sent and how failures are recorded. The objective is reliable task completion, rather than membership of a particular camp.
5. Start from the user's final intent, and allow time for validation
Listen: 32:46–45:05. Wang describes search as a shared intermediate step across many tasks. People ultimately want to complete research or handle an actual matter; searching gathers some of the necessary information. If AI organizes services around that final intent, future products need not merely add a chat entry point to each old app button.
Understanding intent still differs from completing a service. Producing a plan, finding information, receiving consent, invoking a service, and confirming completion cannot be hidden inside a fluent answer. The earlier capability distinction returns: interaction may change before delivery becomes automated.
The later discussion concerns capital and cognition. Wang argues that venture investment has small samples and slow feedback. Capital that demands predictable short term profit can reduce support for uncertain exploration. This describes a tension in an investment mechanism; it does not establish that an entire category of funds has disappeared, nor offer personal investment advice.
Product teams can borrow the lesson about pace. Do not rewrite every assumption after each launch, but do not invoke patience to avoid looking at users. Long horizon beliefs and short experiments can coexist: the former guide exploration, while the latter determine which route deserves more resources.
Exercise: separate three kinds of evidence in an AI proposal
Choose an existing project and write its entire process, from a user's request to confirmed completion. Then fill in the following distinctions:
| Claim | Minimum evidence needed | What it does not yet establish |
|---|---|---|
| Users express needs more easily | Less effort understanding and operating real tasks | Independent execution of every task |
| One step can be automated | Representative inputs, failures, and human takeover records | Reliable operation across an extended workflow |
| A complete workflow can be delivered | State, permissions, recovery, cost, and completion confirmation | Unlimited extension into any industry |
| Early investment retains value | Transferable customer relationships, lawful feedback, and capabilities | Inevitable victory over later entrants |
Add two lines: which cost assumption supports the current economics, and what scope would shrink if that assumption arrived a year late. This is more actionable than arguing only about whether AI still deserves belief.
Source boundary
This reading uses the episode's official introduction and transcript. Industry trajectories, falling costs, and capital assessments are views expressed in 2024; their subsequent fulfillment has not been independently verified. A passage around minute 18 contains a negative that conflicts with its context, so no precise conclusion is drawn from it. Uncertain device prices and model measurements are also omitted. Editorial inferences and the exercise are not presented as the guest's words.
Listening navigation
- 01:28–08:58: Early mobile internet mistakes and differences from AI.
- 09:18–17:25: Three capability layers and their conditions.
- 18:50–23:23: Opportunity sequence and incumbent strengths.
- 23:51–28:32: Founder adaptability and early investment.
- 28:43–34:34: Historical cost forecasts, model choices, and final intent.
- 35:43–45:05: Incumbent limits, capital mechanisms, and feedback cycles.
Related themes: AI industry stages, product economics, startup timing. Further listening: Do not sprint just after starting and Grasp and agent execution environments.